← board

Missing FPC surface found by the differential probe (Eoln, TSeekOrigin, Sorted, IncMonth)

Each is a hard "undefined"/"no such member" error, so none can miscompile — they are missing surface, not wrong behaviour. Grouped because they were found in one sweep and each is small.

The items

  1. Eoln(f) — undefined variable (Eoln). Eof exists. Eoln is how character-at-a-time reading finds the end of a line, so its absence forces a readln-and-index workaround.

  2. TSeekOrigin constants — undefined variable (soFromBeginning). TStream.Seek itself exists; the origin constants (soFromBeginning / soFromCurrent / soFromEnd, and the modern soBeginning / soCurrent / soEnd) do not, so Seek cannot be called portably.

  3. TStringList.Sorted — no such member. Sort (the one-shot method) exists; the Sorted property, which keeps the list ordered across subsequent Adds and switches IndexOf to a binary search, does not. Note the semantic difference from Sort when taking this on: setting Sorted := True changes where later Adds land, and FPC raises on Insert into a sorted list. Also missing alongside it: Duplicates (dupIgnore/dupError/dupAccept), which only has meaning when sorted.

  4. IncMonth — undefined variable. The end-of-month clamp is the whole point and the part worth reading off FPC rather than deriving: IncMonth(EncodeDate(2026,1,31), 1) is 2026-02-28, not an overflow into March.

Append, FileExists, DeleteFile and IsLeapYear were suspected in the same sweep and all exist and behave correctly — the probe cases that seemed to implicate them were actually failing on bug-p-writeln-text-rejects-char. Checked individually before filing.

Method note

Each has a probe case waiting in tools/fpc_diff_probe.sh, tagged known. Whoever implements one should read the expectations off an FPC build rather than from the documentation — IncMonth's clamp and Sorted's effect on Add are both the kind of rule that is easy to state slightly wrong.

Resolution 2026-08-09 (Track B) — measured first; half of it was already done

The ticket's own method note says to read expectations off an FPC build rather than the documentation. Applying that to the LIST as well as the values changed what the work was: compiling each of the four items showed two of them already implemented since this was filed on 2026-08-04.

item measured state done
Eoln undefined variable (Eoln) added
soFromBeginning / soFromCurrent / soFromEnd undefined added as aliases
soBeginning / soCurrent / soEnd already present, Ord 0/1/2 guarded
TStringList.Sorted + Duplicates already present and correct — output matched FPC line for line, including Add landing in order, binary IndexOf, dupIgnore/dupAccept, and Ord(dup*) = 0/1/2 guarded
IncMonth undefined variable added

Eoln (lib/rtl/textfile.pas) reuses the one-byte lookahead Eof already parks in f.Peek, so it is non-destructive and shares a cursor with everything else. FPC-measured, and the CR row is the one worth naming: Eoln answers True at #13 even though TextReadChar hands #13 back as an ordinary character. Both are FPC's behaviour, each measured on its own rather than inferred from the other.

soFrom* are aliases of the existing enum values, not a second enum, so a Seek written either way is the same call.

IncMonth (lib/rtl/sysutils.pas). The clamp is exactly what the ticket warned it would be, and the case that catches a wrong implementation is +2 from the 31st: the clamp applies to the ORIGINAL day against the FINAL month, so IncMonth(2026-01-31, 2) is 2026-03-31, not the 2026-02-28 that clamping month-by-month would give. Leap February, both directions, year rollover and time-of-day preservation (including across a clamp) all measured against FPC.

Regression test: test/lib_fpc_surface_2026_08.pas, in make lib-test — the guards on the two already-working items included, since "found working" is only useful if it stays that way.

Log